The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local helper script (ccsn-isolated) which runs the project's source code using the bun runtime. This is standard behavior for a development-focused verification tool.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process and analyze potentially dangerous command strings (e.g., rm -rf /, git reset --hard) to verify the safety tool's protection logic.
Ingestion points: Test commands are passed as arguments to explain or as part of JSON payloads to the hook command (e.g., in SKILL.md and features/hook-protection.md).
Boundary markers: The skill provides explicit, repeated instructions to the agent that these commands are "analyzer INPUT only" and must "Never execute one in a shell."
Capability inventory: The harness has the ability to execute the CLI via bun, terminate specific processes via kill, and delete the isolated test directory via rm -rf.
Sanitization: The skill relies on the underlying cc-safety-net tool to handle these inputs as data rather than executable instructions.
[SAFE]: The skill implements strong isolation by redirecting critical environment variables (HOME, USERPROFILE, CC_SAFETY_NET_HOME, CC_SAFETY_NET_AUDIT_HOME) to a disposable directory. It also includes instructions for authenticated access to its local GUI and specific cleanup procedures to ensure no persistent state or orphaned processes remain after a run.

verify-ccsn